Side_1_360

When the srTCM operates in colour-aware
mode, arriving packets are considered to be pre-
marked. Then, the marking must be more con-
servative in the colouring of packets. That is,
the re-colouring is only allowed if it results in
a higher drop probability (change green to yel-
low/red or change yellow to red) for the packet.
The algorithm described above is followed when
a green packet arrives. When an arriving packet
is pre-coloured as yellow, only the status of the
Etoken bucket is considered. If the packet has
a size of Bbytes, the packet remains yellow if
token bucket Econtains at least Btokens upon
its arrival. Otherwise, it is re-coloured as red.
A packet pre-coloured as red remains red.

Two token buckets can be used also for mod-
elling the trTCM (Figure 4). Tokens are added to
the token buckets at rates CIR for Cand PIR for
P. In colour-blind mode, a packet of size Bbytes
is coloured as red if token bucket Pcontains less
than Btokens upon its arrival. If token bucket P
contains at least Btokens, it is checked whether
token bucket Calso contains Btokens. If it does,
the packet is coloured as green and Btokens are
removed from both buckets. Otherwise, it is
coloured as yellow and Btokens are removed
only from token bucket P. The colour-aware
mode of operation is similar to the above de-
scription.

As mentioned above the TCM policing is fre-
quently carried out at the boundary of a DiffServ
domain. Boundary nodes could limit traffic car-
ried on behalf of customers to the constraints
specified in the associated Traffic Conditioning
Specifications (see Chapter 4).

2.2 Buffer Management

The basic buffer management scheme is to treat
all packets equally; insert the packets into a
queue upon arrival and take them out of the

queue for transmission on a link. The buffer

management schemes may be operating on dif-

ferent aggregates of traffic flows, e.g. from all

packets on an interface to individual traffic

flows.

Applying Tail-Drop (TD) arriving packets are

dropped only when the queue is full. A problem

with tail drop is that global synchronisation of

TCP sources can occur as multiple TCP sources

reduce their transfer rates almost at the same

time. Then congestion clears and the TCP

sources gradually increase their transmission

rates again until a new congestion situation may

build up. This could result in longer periods dur-

ing which the transmission link is not fully

utilised. That is, oscillations of the link load

could be observed with a poor average utilisa-

tion.

2.2.1 Random Early Detections and

Derivatives

Active queue management mechanisms can be

designed to avoid the synchronisation of TCP

connections. A goal of such an active queue

management mechanism is to make each TCP

connection reduce its sending rate at different

moments. An essential algorithm for this is the

Random Early Detection (RED). When applying

this algorithm packets are discarded randomly

when the buffer is near to congestion. In this

way, TCP connections will lose packets at dif-

ferent time instants, avoiding the synchronisa-

tion between TCP connections. RED supports

congestion avoidance by controlling the average

queue size. During congestion (but before the

queue is filled), the RED scheme marks arriving

packets according to a probabilistic algorithm

which takes into account the average queue size.

The marked packets can be dropped as an early

congestion notification before queues actually

overflow. This will trigger corresponding TCP

Figure 4 Single rate (left) and

two rate (right) Three Colour

Marking

Tc credit

Tp credit

blind

G

Y

R

X X G G Y R

X

R

R

R

R

trTCM

X

Y

Y

Y

R

R

R

R

R

G+Y

Tp credit

PBS

G

Tc credit

CBS

CIR PIR

Tc credit

Te credit

blind

G

Y

R

X X G G Y R

X

G

G

Y

R

srTCM

X

Y

Y

Y

R

R

R

R

R

Y

Te credit

EBS

G

Tc credit

CBS

CIR spillover

CE C P